molecular comparisons by maximum common sub...

MedChemica | 2016

Molecular comparisons by Maximum Common Sub-Structure (MCSS) and application to Matched Molecular

Pair Analysis (MMPA)

MedChemica BIGCHEM Lecture Oct 2016

SlideShare:

MedChemica | 2016

Lecture: •  Brief introduction to MedChemica – MCSS and MMPA at the

heart of pharma knowledge sharing •  Comparing molecules:

–  Why do it? –  What methods, compare contrast? –  What is your experiment?

•  About MCSS: –  Definitions –  A simple use case – depicting the common structure of two molecules –  What tools are out there? –  A brief bit of code in python (with RDkit and OpenEye toolkits)

•  Matched Molecular Pair Analysis: –  What is it? –  MCSS is just the start of the problem….chemical encoding –  Processing the data – generating rules

•  What next? MMPA methodology can be extended to extract pharmacophores

MedChemica | 2016

MedChemica – Enabling Knowledge sharing →Better medicinal chemistry

Full Paper, J.Med.Chem submi.ed

MedChemica | 2016

Recent Media

“Since the big 10 pharma companies spend $70 billion a year collec=vely on research without always have a good return, collabora=ng on research can benefit companies from all the money spent by all companies in the industry. This collabora=on will benefit both AstraZeneca and Roche without divulging confiden=al informa=on” – NASP – 26th June

The ‘Market’ approves of pharma collabora=ng

MedChemica | 2016

Lecture: •  Brief introduction to MedChemica – MCSS and MMPA at the

heart of pharma knowledge sharing •  Comparing molecules:

–  Why do it? –  What methods, compare contrast? –  What is your experiment?

•  About MCSS: –  Definitions –  A simple use case – depicting the common structure of two molecules –  What tools are out there? –  A brief bit of code in python (with RDkit and OpenEye toolkits)

•  Matched Molecular Pair Analysis: –  What is it? –  MCSS is just the start of the problem….chemical encoding –  Processing the data – generating rules

•  What next? MMPA methodology can be extended to extract pharmacophores

MedChemica | 2016

Comparing Molecules •  Chemical information processing is the science of representing molecules in

computers. Hence the fundamental “object” or data structure within a chemical information system is that of the molecule, its atoms (nodes) and its bonds (edge).

Why Compare molecules?

Database searching

Chemical representa=on

and interpreta=on

Structure Ac=vity / Property Rela=onships

Knowledge extrac=on

A word about experimental design: Always consider what question is being asked by the experiment, by considering this we can choose the right technique to use Compute tasks can take days to run, especially comparing molecules

MedChemica | 2016

Maximum Common Sub Structure (or Graph) Method

A/B 36/39 ha = 92% MCS overlap

•  Structures matched by overlapping of the matching atoms (nodes) and bonds (edges)

•  The compute process uses works by traversing atoms and bonds in a graph representation of the molecules in memory

•  The compute problem is NP-complete or NP-hard.

A - CHEMBL1784632 B - CHEMBL2316582

Non-matching heavy atom

MedChemica | 2016

Maximum Common Sub-Structure patterns versus Fingerprints

MCSS Process •  Convert each molecule into a

graph representation in memory. Traverse each molecule finding atoms (nodes) and bonds (edges) that are the same, loop many many times building the largest set of edge linked nodes.

•  Larger memory •  Slow (many algorithm set

a time-out) •  Result is concise atom/

bond structure, so difference in chemistry can be captured

Fingerprints Process •  Convert each molecule into a ‘bit

string (0s and 1s) representing parts of the molecule – thus each molecule is a ‘number’ in a computer – thus comparison between two molecules is easy.

•  Low memory •  Very Fast •  Result is a numerical

difference between the molecules – no exact chemical structural difference

Experimental design is important – how much time and compute resource do you have?

Maggiora, G; Vogt, M.; Stumpfe, D.; Barorath, J; Molecular Similarity in Medicinal Chemistry, J.Med.Chem.2013,3186.

MedChemica | 2016

Graphs Applied to Molecules h.ps://en.wikipedia.org/wiki/Graph_theory

A Graph (G) is a set of Nodes and Vertices G = (N, V) Each Node has relationship between other nodes by the connections made by Vertices Each Node is connected to every other node by the Vertices and other Nodes…. ….Node 6 is 3 Vertices from Node 2 but there are two paths to get there, via 3 or 5….

MedChemica | 2016

Molecules are suited to Graphs

SMILES Cc1ccc(cc1)C(c2ccccc2)c3ccccc3 ‘parse’ into Molecular Graph For example: OpenEye Toolkit >>> from openeye.oechem import OEGraphMol, OESmilesToMol >>> mol = OEGraphMol() >>> OESmilesToMol(mol, "Cc1ccc(cc1)C(c2ccccc2)c3ccccc3") >>> for atom in mol.GetAtoms(): ... print ('Atom Idx :{0} Atomic Num: {1}'.format(atom.GetIdx(), atom.GetAtomicNum()) Atom Idx :0 Atomic Num: 6 Atom Idx :1 Atomic Num: 6 Atom Idx :2 Atomic Num: 6…. ... for bond in mol.GetBonds(): print (bond.GetOrder())

3

2 1 0

4 5

7

6

8

9 10

11

12 13 14 15

16

17 18

19

Nodes are atoms 109 atom ‘types’ – known bonding behaviour

Vertices are bonds single, double, triple……..

Molecular Graphs are the corner stone for chem-infomatics processing

WARNING – Do Not be tempted to manipulate SMILES strings with text searches or replace func=ons

MedChemica | 2016

Types of MCSS – lets look at some molecules MCES – Maximum Common Edge-‐Induced Substructure – as many matching chemical bonds

cMCES – connected MCES The largest matching fragment of both molecules is itself connected structure

dMCES – disconnected MCES Mul=ple matching parts of the molecule

SMILES Cc1ccc(cc1)C(c2ccccc2)c3ccccc3 Cc1ccc(cc1)N(c2ccccc2)c3ccccc3 cMCES [C][c]1[c][c][c]([c][c]1) dMCES [C][c]1[c][c][c]([c][c]1).[c]2[c][c][c][c][c]2.[c]3[c][c][c][c][c]3

MedChemica | 2016

The Travelling Sales Man Problem (and other NP-hard problems)

Comparing Graphs to find Maximum Common Sub-Graph is compute intensive Described as NP

- Non-deterministic polynomial time - not sure how many iterations over the graphs it will take - Much harder and longer the more complex the graph (lots of atoms and bonds)

h.ps://en.wikipedia.org/wiki/Travelling_Salesman_(2012_film)

What is the shortest possible route for a traveling salesman seeking to visit each city on a list exactly once and return to his city of origin? It has defied solu=on to this day

MedChemica | 2016

MCSS - References

•  Maggiora, G; Vogt, M.; Stumpfe, D.; Barorath, J; Molecular Similarity in Medicinal Chemistry, J.Med.Chem.2013,3186.

•  Cao, Y.; Jiang, T.; Girke, T. A Maximum common substructure-based algorithm for searching and predicting drug-like compounds. Bioinfomatics, 2008, 366.

•  Duesbury, E.; Holliday, J.; Willet, P. Maximum Common Substructure-Based Data Fusion in Similarity Searching, J.Chem.Inf.Model 2015, 222.

•  Hariharan, R.; Janakiraman, A.; Nilakantan, R.; Singh, B.; Varghese, S.; Landrum, G.; Schuffenhauer, A. MultiMCS: A Fast Algorithm for the Maximum Common Substructure Problem on Multiple Molecules, J.Chem.Inf.Model 2011, 786.

MedChemica | 2016

A difficult connected MCSS – can you spot it?

CHEMBL1779022 CHEMBL2146793

14

Paired by MCSS

Hard compute – 54 seconds for just this pair on one CPU MCSS is compute intensive – lets use this example 400 antibiotic macrocycles - ALL pair comparison to find Structure Activity Relationships (SAR) Roughly how long?

(400 x 400 x 54 secs) / 2 / 3600 / 24 = 50 days for 1 CPU ….we are going to need a 50 CPUs to run in a day.

MedChemica | 2016


15

Paired by MCSS

This is a useful case study: •  MCSS is useful as a sub-substructure that matches both molecules •  Using a depictions tool we can visualise the difference between them •  For compound design / optimisation this is very useful

A difficult MCSS – can you spot it?

MedChemica | 2016

MCSS on Macrocycles


16

Paired by MCSS Common substructure and change is clear

Hard to re-‐draw by a human – 30 minutes!

MedChemica | 2016

MCSS on Macrocycles


17

Graph processing algorithm has worked as coded, but did not pick out the difference between cyclised and open chain.

So we need to discover new drugs; as molecules become more complex (anti-biotic macrocycles) using a computer to help analyse and design new molecules becomes hard

MedChemica | 2016

Molecule Fragment and Index Method


a a

Break all single rotatable bonds and group on common ‘core’ parts

•  Semantically the same matched pair as MCSS but syntactically different in that ‘points of modification’ are not the same

MedChemica | 2016

Frag / Index single bond SMARTS From the Hussain and Rea publication Hussain, J., & Rea, C. (2010). "Computationally efficient algorithm to identify matched molecular pairs (MMPs) in large data sets." Journal of chemical information and modeling, 50(3), 339-348.

A SMARTS pattern is required to identify the single bonds to break Original [*]!@!=[*] (any atoms, any bond not in ring or double – what about triple bonds?) Amended in paper (see ref 24) [#6+0;!$(*=,#[!#6])]!@!=!#[*] (carbon atom to any atom [excluding carbon that also has a double or triple] and no triple bond]) Effectively this removes amide bonds as disconnections MedChemica variation [#6+0;!$(*=,#[!#6])]!@!=!#[*;!$([H])] Carbon atom to any atom excluding hydrogen – stop disconnection to H as part of a explicit chiral centre SMARTS pattern can be set in config file or from the command line for direct .smi file processing

Bonds broken by pattern

not broken by pattern

MedChemica | 2016

Frag / Index and MCSS pair finding Consider the following compounds

Strict heavy atom c-MCSS

FI and MCSS have found the ‘same’ pattern for the matched pairs via the same core. FI requires additional compute to find the true MCSS

Three ways of Frag / Index finding matches (dependent on settings) Num. of Trans = 3 x num of Environment atoms

The bond to H may not be broken but will be added by the H-additions step of the process


MedChemica | 2016

Frag And Index – useful for disconnected MCSS

A B

A and B are matched by the same context groups (for MCSS these are disconnected graphs so are not currently found) Context can be captured on both ends in the SMIRKS

Both acyclic to acyclic and acyclic to cyclic changes are captured

Subtle changes in ring isomers are found in the ‘centre’ of molecules

A – CHEMBL100461 B –CHEMBL103900

A – CHEMBL10085 B – CHEMBL10327

For large molecule whole side chain modification can be captured


Molecule Fragment and Index Method: ‘Linker’ changes found by double cuts

MedChemica | 2016

MCSS – subtle differences and ring changes Maximum Common Sub Structure (or Graph) Method

For molecules A and B of the same size (by heavy atom count) that are smaller (<30) transformations are found for positional isomer switches (these can be missed by F/I)

Subtle changes to smaller groups around rings are captured for all examples.

A B



Acyclic to cyclic changes are captured well by MCSS method

A – CHEMBL156639 B - CHEMBL2387702

MedChemica | 2016

Does the comparison method really matter?

Using only one technique will miss between 12% and 56% of pairings

23

Pairings Pairings number of compounds common FI only MCSS only total FI only % common % MCSS only %

VEGF 4466 14631 17172 14823 46626 37 31 32 Dopamine Transporter 1470 4480 8930 3497 16907 53 26 21

GABAA 848 2500 1722 4205 8427 20 30 50

D2 human 3873 12995 13811 13098 39904 35 33 33

D2 rat 1807 5408 6595 7346 19349 34 28 38 Acetylcholine esterase 383 536 725 1434 2695 27 20 53 Monoamine oxidase 264 653 1156 246 2055 56 32 12

min 20 20 12

max 56 33 53

FI MCSS

co

mm

on

MedChemica | 2016

MCPairsv1.x Settings for FI and MCSS methods

A/B 36/39 ha ratio = 0.92 MCS overlap


Non-matching heavy atom

Set MCSSHAcutoff in configs smaller overlap <0.9 higher overlap >0.9

a a

modi / core ha ratio = 6/30 ratioFI = 0.2

modi / core ha ratio = 9/30 ratioFI = 0.3

Set FIratio in configs larger inclusion >0.3 lower inclusion <0.3

MedChemica | 2016

OpenEye ToolKit MCSS #!/usr/bin/env python from __future__ import print_function from openeye.oechem import * pattern = OEGraphMol() target = OEGraphMol() OESmilesToMol(pattern, "c1cc(O)c(O)cc1CCN") OESmilesToMol(target, "c1c(O)c(O)c(Cl)cc1CCCBr") atomexpr = OEExprOpts_DefaultAtoms bondexpr = OEExprOpts_DefaultBonds # create maximum common substructure object mcss = OEMCSSearch(pattern, atomexpr, bondexpr, OEMCSType_Exhaustive) # set scoring function mcss.SetMCSFunc(OEMCSMaxAtoms()) # ignore matches smaller than 6 atoms mcss.SetMinAtoms(6) unique = True # loop over matches for count, match in enumerate(mcss.Match(target, unique)): print ("Match %d:" % (count + 1)) print ("pattern atoms:", end=" ") for ma in match.GetAtoms(): print (ma.pattern.GetIdx(), end=" ") print ("\ntarget atoms: ", end=" ") for ma in match.GetAtoms(): print (ma.target.GetIdx(), end=" ") # create match subgraph m = OEGraphMol() OESubsetMol(m, match, True) print ("\nmatch smiles =", OEMolToSmiles(m))

h.p://docs.eyesopen.com/toolkits/python/oechemtk/pa.ernmatch.html

MedChemica | 2016

Rdkit MCSS – Python >>> from rdkit.Chem import rdFMCS >>> mol1 = Chem.MolFromSmiles("O=C(NCc1cc(OC)c(O)cc1)CCCC/C=C/C(C)C") >>> mol2 = Chem.MolFromSmiles("CC(C)CCCCCC(=O)NCC1=CC(=C(C=C1)O)OC") >>> mol3 = Chem.MolFromSmiles("c1(C=O)cc(OC)c(O)cc1") >>> mols = [mol1,mol2,mol3] >>> res=rdFMCS.FindMCS(mols) >>> res <rdkit.Chem.rdFMCS.MCSResult object at 0x...> >>> res.numAtoms 10 >>> res.numBonds 10 >>> res.smartsString '[#6]1(-[#6]):[#6]:[#6](-[#8]-[#6]):[#6](:[#6]:[#6]:1)-[#8]' >>> res.canceled False

h.p://www.rdkit.org/docs/GemngStartedInPython.html#maximum-‐common-‐substructure

MedChemica | 2016

About Clean Code

Compute code is written once and read a thousand times Do:

write Unit-Tests (especially with processing molecules) write functional tests write looooooong function, variable names write functions that DO ONE THING! write modular code – DRY (Do-not Repeat Yourself)

Clean Code: A Handbook of Agile So7ware Cra7smanship (Robert C. MarBn)

Working EffecBvely with Legacy Code (Micheal C. Feathers)

MedChemica | 2016

MedChemica | 2016

Matched Molecular Pair Analysis

An example of using MCSS (and FI) to compare molecules and extract Medicinal chemistry knowledge – Example it to show that comparing

molecules it just the start of a process – What experiment are we doing? – How do we process the chemistry output

and the data? – What things can go wrong? – Chirality!

MedChemica | 2016

•  Matched Molecular Pairs – Molecules that differ only by a particular, well-defined structural transformation

•  Transformation with environment capture –MMPs can be recorded as transformations from A B

•  Environment is essential to understand chemistry

Statistical analysis •  Learn what effect the transformation has had on ADMET properties in

the past

Griffen, E. et al. Matched Molecular Pairs as a Medicinal Chemistry Tool. Journal of Medicinal Chemistry. 2011, 54(22), pp.7739-‐7750.

Matched Molecular Pair Analysis uses MCSS methods….

Δ Data A-B 1

2

2

3

3

3

4

4

4

1 2

2 3

3

3 4

4

4 A B

MedChemica | 2016

Identify and group matching SMIRKS

Calculate statistical parameters for each unique SMIRKS (n, median, sd, se, n_up/n_down)

Is n ≥ 6?

Not enough data: ignore transformation

Is the |median| ≤ 0.05 and the intercentile range (10-90%) ≤ 0.3?

Perform two-tailed binomial test on the transformation to determine the

significance of the up/ down frequency

transformation is classified as ‘neutral’

Transformation classified as ‘NED’ (No Effect Determined)

Transformation classified as ‘increase’ or ‘decrease’

depending on which direction the property is changing

passfail

yesno

yesno

Rule selection

0 +ve -ve

Median data difference

Neutral Increase Decrease

NED

MedChemica | 2016

Environment really matters HMe: •  Median Δlog(Solubility) •  225 different

environments

2.5log

1.5log

HMe: •  Median Δlog(Clint)

Human microsomal clearance

•  278 different environments

MedChemica | 2016

More environment = right detail HMe Solubility: •  225 different environments

MedChemica | 2016

HF What effect on Clearance? •  Median Δlog(Clint) Human microsomal clearance •  37 different environments

2 fold improvement 2 fold worse

Increase clearance

decrease clearance

MedChemica | 2016

Chiral Test Set example

MedChemica | 2016

Chirality MCPairs --chiralON

MedChemica | 2016

How to ‘store’ chemical information?

MCSS MMPA SMIRKS Storing reactions as canonical SMIRKS – an example of storing chemical information

MedChemica | 2016

Canonicalised SMIRKS 1 – Its about Symmetry

4-‐Atom rule

Both SMIRKS are VALID and would transform a molecule correctly However we want to consistently produce a single SMIRKS for all of these It is 50:50 how an unchecked algorithm will number around the ring Therefore we force the numbering in the reactant side by numbering 1,2,3,4..n Then use symmetry checking to change the numbers to a consistent pattern Thus SMIRKS_B is produced

4 3 2

1

4 3 2

1

SMIRKS_A [C]([H])([H])([H])[O][c:1]1[c:2]([H])[c:3]([H])[c:4][c:5]([H])[c:6]1([H])

>>[c:5]1([H])[c:6]([H])[c:1]([c:2]([H])[c:3]([H])[c:4]1)[Br] SMIRKS_B [C]([H])([H])([H])[O][c:1]1[c:2]([H])[c:3]([H])[c:4][c:5]([H])[c:6]1([H])

>>[c:3]1([H])[c:2]([H])[c:1]([c:6]([H])[c:5]([H])[c:4]1)[Br]

6 5

6 5

4

3 2 1

4

3 2 1

6 5

6 5

MedChemica | 2016

Canonicalised SMIRKS 2


1-‐Atom rule [H][O:1]>>[C]([H])([H])([H])[C]([H])([H])[O:1]

Highly specific explicit H

Key mapped atom With 1 atom rule this is simple – 1 becomes 1

2-‐Atom rule [H][O:1][c:2]>>[c:2][O:1][C]([H])([H])[C]([H])([H])([H])

Mapped atoms Absolutely key – note the ethyl is right most in SMIRKS

2

1 3

3 4 4

Orange – atom env radius Blue -‐ atom map index

MedChemica | 2016



3-‐Atom rule [H][O:1][c:2]([c:3])[n:4]>>[c:3][c:2]([n:4])[O:1][C]([H])([H])[C]([H])([H])([H])

Highly specific explicit H

Key mapped atom With 3 atom rule environment is more complex

4-‐Atom rule [H][O:1][c:2]1[c:3]([c:4][o:5][n:6]1)[C:7]([H])([H])>> [C]([H])([H])([H])[C]([H])([H])[O:1][c:2]1[c:3]([c:4][o:5][n:6]1)[C:7]([H])([H])

Note Mapped atoms run left to right 1,2,3,4…n Absolutely key – note the ethyl is LEFT most in SMIRKS Rule change depending on rule size, environment and symmetry

2

1 3

3 4 4

4 Orange – atom env radius Blue -‐ atom map index

4

3 2

1

4 3

2 1

MedChemica | 2016



3-‐Atom rule [H][O:1][c:2]([c:3])[n:4]>>[c:3][c:2]([n:4])[O:1][C]([H])([H])[C]([H])([H])([H])

Key mapped atom Explicit Hydrogen are on the RHS of product side Thus the mappings become complex and NOT in order 1,2,3,n

While this SMIRKS is VALID and would transform a molecule correctly IF we search for the this SMIRKS there is NO MATCH

Reverse SMIRKS [c:3][c:2]([n:4])[O:1][C]([H])([H])[C]([H])([H])([H])>>[H][O:1][c:2]([c:3])[n:4]

Actual SMIRKS [c:1][c:2]([n:3])[O:4][C]([H])([H])[C]([H])([H])([H])>>[H][O:4][c:2]([c:1])[n:3]]

Due to Explicit H, Symmetry and Canonicalisation - Map Index must be treated with care – directly swapping the string may not match

NOTE – The same applies to fragment SMILES

4 3

2 1

4 3

2 1 Orange – atom env radius

Blue -‐ atom map index

MedChemica | 2016

Take home messages •  Molecular Graphs are the corner stone of chem-infomatics

•  Allow traversal of atoms and bonds

•  Allow graph theory techniques to be applied to comparing molecules

•  Maximum Common Sub-Structure is found by NP-complete graph techniques and is compute intensive

•  The Fragment and Index method can find MCSS in many circumstances but not all

•  Further compute processes are required to deal with chirality and encoding of the common parts of molecules and the difference between molecules

•  Matched Molecular Pair analysis is a data analysis technique, based on MCSS, to extract chemical design knowledge

MedChemica | 2016

Appendix

MedChemica | 2016

IP security

•  Contributing company structures are NOT shared •  Contributing company identifiers are NOT shared •  Each contributing company receives it’s OWN

custom GRD copy to enable drill back to OWN example data

•  Minimum of n=6 example pairs required for inclusion in the GRD

•  Enforced limits of shared substructure sizes •  Option for “time-blocking” sharing of data to

withhold data < 6 months old for IP submission

MedChemica | 2016

MCPairs Platform

•  Extract rules using Advanced Matched Molecular Pair Analysis •  Knowledge is captured as transformations

–  divorced from structures => sharable

Measured Data

rule finder Exploitable

Knowledge

MC Expert Enumerator

System

Problem molecule

Solution molecules

Pharmacophores& toxophores

SMARTS matching

Alerts Virtual screening Library design

Protect the IP jewels

MedChemica | 2016

Barriers Broken to Sharing Knowledge

Data Integrity and

curation Knowledge extraction algorithms

Consortium building to

share knowledge Into the minds of

chemists

✓

MedChemica | 2016


Data Integrity and




chemists

✓ ✓

MedChemica | 2016

Merging knowledge •  Use the transforms that

are robust in both companies to calibrate assays.

•  Once the assays are calibrated against each other the transform data can be combined to build support in poorly exemplified transforms

•  Methodology precedented in other fields

Calibrate Robust

Robust

Weak

Weak

Discover Novel

Pharma 1

Pharma 2

MedChemica | 2016

Merging Datasets •  Datasets are standardized by comparison of

transformations shared by contributing companies

•  Transformations are examined at the “pair example” level

•  Minimum of 6 transformations, each with a minimum of 6 pairs (42 compounds bare minimum) required to standardise

•  “calibration factors” extracted to standardize the datasets to a common value – mean of calibration factors 0.94, typical range 0.8-1.2.

•  Datasets with too few common transformations have standard compound measurements shared for calibration.

“Blinded” source of transforma=ons

MedChemica | 2016

Current Knowledge sets – GRDv3 Numbers of statistically valid transforms

Grouped Datasets Number of Rules

logD7.4 153449

Merged solubility 46655

In vitro microsomal clearance: Human, rat ,mouse, cyno, dog 88423

In vitro hepatocyte clearance : Human, rat ,mouse, cyno, dog 26627

MCDK permeability A-B / B – A efflux 1852

Cytochrome P450 inhibition: 2C9, 2D6 , 3A4 , 2C19 , 1A2 40605

Cardiac ion channels NaV 1.5 , hERG ion channel inhibition 15636

Glutathione Stability 116

plasma protein or albumin binding Human, rat ,mouse, cyno, dog 64622

MedChemica | 2016

Pharma 1 100k rules

Pharma 2 92k rules

Pharma 3 37k rules

5.8k rules in common (pre-merge) ~ 2%

New Rules 88k ~26% of total

Merge

Combining data yields brand new rules Gains: 300 -‐ 900%

Merging knowledge – GRDv1

MedChemica | 2016

Key findings:

•  Secure sharing of large scale ADMET knowledge between large Pharma is possible

•  The collaboration generated great synergy •  Many findings are highly significant. •  MMP is a great tool for idea generation. •  The rules have been used in drug-discovery projects

and generated meaningful results •  MMPA methodology can be extended to extract

pharmacophores

MedChemica | 2016


Data Integrity and




chemists

✓ ✓

✓

MedChemica | 2016

Integrating knowledge exploitation

..many possible connections

..high Stability and flexibility**

Rule Database

MCExpert

API server

** api.medchemica.com has delivered 18 months of uptime and drives SaltTraX (Elixir) Approx 50 chemists use the tool

RESTful API

Chemistry Shape and electrostaHcs

MedChemica | 2016

Matched Molecular Pair

data A data B

data C data D

data E data F

Chemical Transformations

Δ data A B

Δ data C D

Δ data E F

Chemical Transformations

Δ data A B

Δ data C D

Δ data E F

Δ data G H

Δ data I J

Δ data K L

Matched Molecular Pair Analysis (MMPA) enables SAR sharing

Without sharing underlying structures and data

Grand Rule

Database

Enumeration

Rate-My-Idea

GRD-Browser

ChEMBL Tox database Toxophores MC-

Biophore

MedChemica | 2016

Collaborators and Users - experience

MedChemica | 2016

A Collaboration of the willing

Craig Bruce OE John Cumming Roche David Cosgrove C4XD Andy Grant★

Martin Harrison Elixir Huw Jones Base360 Al Rabow Consulting David Riley AZ Graeme Robb AZ Attilla Ting AZ Howard Tucker retired Dan Warner Myjar Steve St-Galley Syngenta David Wood JDR Lauren Reid MedChemica Shane Monague MedChemica Jessica Stacey MedChemica

Andy Barker Consulting Pat Barton AZ Andy Davis AZ Andrew Griffin Elixir Phil Jewsbury AZ Mike Snowden AZ Peter Sjo AZ Martin Packer AZ Manos Perros Entasis Therapeutics Nick Tomkinson AZ Martin Stahl Roche Jerome Hert Roche Martin Blapp Roche Torsten Schindler Roche Paula Petrone Roche Christian Kramer Roche Jeff Blaney Genentech Hao Zheng Genentech Slaton Lipscomb Genentech Alberto Gobbi Genentech

MedChemica | 2016

MedChemica | 2016

ACS Philadelphia 2016

- Fix hERG problem whilst maintaining potency

Waring et al, Med. Chem. Commun., (2011), 2, 775

Glucokinase Activators

MMPA ∆pEC50: -0.1 ∆logD: -0.6 ∆hERG pIC50 :-0.5

n=33 n=32 n=22

MMPA ∆pEC50: +0.3 ∆logD: +0.3 ∆hERG pIC50 :-0.3

n=20 n=23 n=19


n=27 n=27 n=7

MedChemica | 2016


A Less Simple Example Increase logD and gain solubility

Property NumberofObserva2ons

Direc2on MeanChange Probability

logD 8 Increase 1.2 100%

Log(Solubility) 14 Increase 1.4 92%

Whatistheeffectonlipophilicityandsolubility?Rochedataisinconclusive!(2pairsforlogD,1pairforsolubility)

logD=2.65KineMcsolubility=84µg/mlIC50SST5=0.8µM

logD=3.63KineMcsolubility=>452µg/mlIC50SST5=0.19µM

Ques2on:

AvailableSta2s2cs:

RocheExample:

MedChemica | 2016


Solving a tBu metabolism issue

Benchmarkcompound

Predictedtooffermostimprovementinmicrosomalstability(inatleast1species/assay)

R2R1

tBu Me Et iPr

99392

1664

78410

53550

99288

78515

4135

98327

92372

24247

35128

2462

60395

39445

321

2027

5789

5489

•  Data shown are Clint for HLM and MLM (top and bottom, respectively)

R1 R2R1tBuRoger Butlin Rebecca Newton Allan Jordan

MedChemica | 2016

Cathepsin K – Di-methoxy surprise – Man and Machine

pIC50 7.95 LogD 0.67 HLM <2.0 Solubility 280μM DTM ~1.0 mg/kg UID Potent Too polar / Renal Cl

PDB -‐ 97% of structures Crawford, J.J.; Dosse.er, A.G J Med Chem. 2012, 55, 8827. Dosse.er, A. G. Bioorg. Med. Chem. 2010, 4405 Lewis et al, J Comput Aided Mol Des, 2009, 23, 97–103

pIC50 8.2 LogD 2.8 HLM <1.0 Solubility >1400μM DTM 0.01 mg/kg UID High F% / stability maximised

Increase in LogP, Properties improved

Solubility ΔpIC50 - 0.1 ΔLogD +1.4 ΔpSol +1.2 ΔHLM + 0.25

No renal Cl low F%

ΔpIC50 +0.1 ΔLogD - 0.7 ΔpSol ~0.0 ΔHLM - 0.25

High F% rat/Dog Electrosta=c poten=al minima between oxygens

Approx like N from 5-‐het, new compound can not form a quinoline

Incr. selec=vity

ΔpIC50 +0.1 ΔLogD - 0.7 ΔpSol ~0.0 ΔHLM - 0.25

High F% rat/Dog

MedChemica | 2016

- Fix hERG problem whilst maintaining potency

Waring et al, Med. Chem. Commun., (2011), 2, 775

Glucokinase Activators


n=33 n=32 n=22

MMPA ∆pEC50: +0.3 ∆logD: +0.3 ∆hERG pIC50 :-0.3

n=20 n=23 n=19


n=27 n=27 n=7

MedChemica | 2016

Knowledge Based Design – MPO –  Novel more efficient core required, improve hERG for CD –  CNS penetration, good potency and deliver tool for in vivo testing

McCoull, Dossetter et al, Med. Chem. Commun., (2013), 4, 456

ΔpIC50 -0.4 ΔlogD -1.8 ΔhERG pIC50 +0.4

Ghrelin Inverse agonists

~

MMPA Cores

pIC50 9.9 logD 5.0 hERG pIC5 5.0 LLE 4.9 very potent very lipophilic

ΔpIC50 +0.9 ΔlogD +0.2 ΔhERG pIC50 -0.3

pIC50 8.2 logD 1.3 hERG pIC50 4.4 LLE 6.9

ΔpIC50 -2.2 ΔlogD -2.2 ΔhERG pIC50 -0.7

100 compounds

made

LLE = lipophilic ligand efficiency: LLE=pIC50-logD

LLE 6.4

LLE 6.9

MedChemica | 2016

Early successes From GRDv1 May 2014

62

J. Med. Chem., 2015, 58 (23), pp 9309–9333 DOI: 10.1021/acs.jmedchem.5b01312

MedChemica | 2016

Comparison of Merck in-house MMPA with SALTMinerTM

Structure:

ADMET Issue: hERG Lead A2A receptor antagonist compound in Merck Parkinson's project

138 suggestion molecules with predicted improvement in hERG

binding

How many match the results of Merck?

•  Also shows potent binding to the hERG ion channel

•  Deng et al performed in-house MMPA on hERG binding compound data and have published 18 resulting fluorobenzene transformations, which they have synthesized and tested for hERG activity

Deng et al, Bioorg. & Med Chem Let (2015), doi: h.p://dx.doi.org/10.1016/j.bmcl.2015.05.036

MedChemica | 2016

R group:

Measured hERG pIC50 change

-‐1.187 -‐1.149 -‐1.038 -‐1.215 -‐1.157 -‐0.149 -‐1.487 -‐1.133

GRD median historic pIC50 change

0 -‐0.171 -‐0.1 -‐0.283 -‐0.219 -‐0.318 -‐0.159 -‐0.103

Results: 8 out of the 18 fluorobenzene transformations produced by Merck were also suggested by MCExpert to decrease hERG binding:

Searching the GRD for transformations that increase hERG there were none that matched the remaining 10 of 18 transformations in the paper.

MCExpert also suggested an additional 50 fluorobenzene replacements to decrease hERG binding NOT mentioned in the publication.

MedChemica | 2016

A Less Simple Example Increase logD and gain solubility

Property Number of ObservaBons

DirecBon Mean Change Probability

logD 8 Increase 1.2 100%

Log(Solubility) 14 Increase 1.4 92%

What is the effect on lipophilicity and solubility? Roche data is inconclusive! (2 pairs for logD, 1 pair for solubility)

logD = 2.65 Kine=c solubility = 84 µg/ml IC50 SST5 = 0.8 µM

logD = 3.63 Kine=c solubility = >452 µg/ml IC50 SST5 = 0.19 µM

QuesBon:

Available StaBsBcs:

Roche Example:

MedChemica | 2016

Solving a tBu metabolism issue

Benchmark compound

Predicted to offer most improvement in microsomal stability (in at least 1 species / assay)

R2 R1

tBu Me Et iPr

99 392

16 64

78 410

53 550

99 288

78 515

41 35

98 327

92 372

24 247

35 128

24 62

60 395

39 445

3 21

20 27

57 89

54 89

•  Data shown are Clint for HLM and MLM (top and bottom, respectively)

R1 R2 R1 tBu Roger Butlin Rebecca Newton Allan Jordan

MedChemica | 2016

The application helped lead optimization in project

67

•  193 compounds •  Enumerated

Objective: improve metabolic stability

MMP Enumeration

Calculated Property Docking

8 compounds synthesized

MedChemica | 2016


Data Integrity and




chemists

MedChemica | 2016

Data Integrity and Curation

Structural •  Extensive standards for

inclusion of mixtures, chiral compounds, salt forms

•  Tautomer and charge state canonicalisation client side

•  Automated validation of structures run client side = “clean” comparable structures submitted to pair finding

Measured Data •  Assay protocols reviewed

prior to merging •  Precise documentation

on unit definitions and data reporting standards

•  Option to share standard compound measured values

•  Automated extensive data validation checks prior to merging data

“client side” = behind Pharma firewall

MedChemica | 2016

Calibrating Assays

•  Sets of transformations can be calibrated against each other as we are comparing Δ values in assays not absolute values

•  Assays are usually linearly displaced against each other •  Data analysis equivalent of FEP

Compound A Compound B Compound C Compound D

Transformation 1 Transformation 2

pIC50, log(Clint), pSol etc

Assay 1 Assay 2

ΔT1 ΔT2

ΔT1’= ΔT1 ΔT2’= ΔT2

ΔT1’ ΔT2’

Assay 2 more sensitive than Assay 1

Assay 1 Δ

Assay 2 Δ

Assay 2 less sensitive than Assay 1

T1

T2

MedChemica | 2016

Pharmacophores and Toxophores by extended analysis from the MMPA

Pharmacophores BigData Stats Matched

Pairs Finding

Public and in-house potency

data

MedChemica | 2016

Target Number of compounds

Number of compound

pairs

Number of Fragments

Number of Pharmacophore dyads ayer filtering

R2 RMSEP ROC odds_ra=o (geomean)

Acetylcholine esterase - human 383 27755 44 10 0.43 1.57 0.80 4 β 1 adrenergic receptor 505 145447 276 313 0.64 0.70 0.96 833 Androgen receptor 1064 113163 186 46 0.47 0.77 0.86 140 CB1 canabinnoid receptor 1104 88091 165 90 0.61 1.02 0.87 96 CB2 canabinnoid receptor 1112 82130 194 158 0.19 0.85 0.64 5.7 Dopamine D2 receptor - human 3873 230962 483 602 0.42 0.88 0.69 110 Dopamine D2 receptor - rat 1807 118736 267 377 0.29 0.85 0.78 125 Dopamine Transporter 1470 106969 282 336 0.58 0.73 0.88 141 GABA A receptor 848 39494 106 167 0.70 0.76 0.97 560 hERG ion channel 4189 242261 392 76 0.61 0.96 0.92 55 5HT2a receptor 642 50870 197 267 0.61 0.59 0.83 600 Monoamine oxidase 264 15439 44 11 0.12 1.25 0.48 181 Muscarinic acetylcholine receptor M1 628 48200 97 510 0.62 0.94 0.89 48

µ opioid receptor 1128 37184 33 11 0.69 1.30 0.87 81

Critical safety target analysis

•  Build models using 10-fold cross validated PLS •  Assess using ROC / BEDROC, R2 vs 100 fold y-scrambled R2 and geomean odds ratio

72

Public Data

Find Matched Pairs

Pharmacophores Find

Pharmacophore dyads

Find Potent Fragments

J. Bowes, et al Nat. Rev. Drug Discov., vol. 11, no. 12, pp. 909–922, Nov. 2012

MedChemica | 2016

Novartis Predictions From Our Model Domain of Applicability….

Actual: 8.4[1] Predicted: 7.5

73


1. J MedChem(2016), Bold et al. 2. MedChem Lett (2016), Mainolfi et al.


Actual: 9.0[2] Predicted: Out of Domain

MedChemica | 2016

MCBiophore GUI screenshot

Assay Image Mean_with Mean_without PLS_coeff Path SMARTS1 SMARTS2 n_examples odds_ratio2

VEGFR 8.3 6.4 0.71 [c]c[c][c] Cc1ccc[c]c1[n] [c]/C(=N/O)/C 18 259.83

VEGFR 8.2 6.4 0.17 [c] [c]c1cc(cc[c]1)/C(=N/OC)/CCc1ccc[c]c1[n] 17 257.60

VEGFR 8.1 6.4 -0.01 [CH3] [cH]c([cH])/C(=N/OC)/C[c]/C(=N/OCC)/C 8 20.21

VEGFR 8.1 6.4 -0.01 [CH3] [c]c1cc(cc[c]1)/C(=N/OC)/C[c]/C(=N/OCC)/C 8 151.2

Detailed results in

excel

MedChemica | 2016

Mining transform sets to find influential fragments

Identify the ‘Z’ fragments associated with a significant number of potency increasing changes – irrespective of what they are replaced with ‘Z’ is ‘worse than anything you replace it with’

Fragment A Fragment B Change in binding

measurement

Public Data

Find Matched

Pairs


+2.7

+3.2

+0.6

+0.6

Identify the ‘A’ fragments associated with a significant number of potency decreasing changes – irrespective of what they are replaced with ‘A’ is ‘better than anything you replace it with’

A

+2.1 +2.2 +1.4

+0.4

+1.8

Z

pKi/ pIC50

Compounds with destructive fragment

Compounds with constructive fragments

Generate Pharmacophore dyads by permuta=ng all the fragments with the shortest path between them

MedChemica | 2016

Toxophores - Detailed, specific & transparent

76

Dopamine D2 receptor human Actual: 9.5

Predicted: 9.1 Mean with: 8.0 Mean without: 6.6 Odds Ratio: 340

Dopamine Transporter Actual: 9.1


GABA-A Actual: 9.0


β1 adrenergic receptor Actual: 7.8 Predicted: 7.7 Mean with: 6.5 Mean without: 5.7 Odds Ratio: 1501


Matched Pairs

Finding

Find Pharmacophore

Dyads

Public and in-house potency

data

MedChemica | 2016

Prediction of unseen new molecules The acid test…

•  Vascular endothelial growth factor receptor 2 tyrosine kinase (KDR)

•  Inhibitors have oncology and ophthalmic indications

•  Large dataset in CHEMBL

•  10 fold cross validated PLS model

•  Selected model by minimised RMSEP

77

Compounds 4466 Matched Pairs 288100 Fragments 678 Pharmacophore dyads 787 RMSEP 0.8 R2 0.64 Y-scrambled R2 0.0 ROC 0.95 Geomean odds ratio 80

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●●

●●

●

●

●

●

●

●

●

● ●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●

●

●

●●●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

● ●

●●

●

●●

●

●

●●

●

●

●

● ●●●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●●

●

●

●

●●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●●

●●

●

●

●

●

● ●●

●●●

●

●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●●●●

●

●●

●●●●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

● ●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●●●

●

●●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●●

●●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●●

●

●

●

●

●●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●

●

●

●

●

●

●●

●●

●

●

●

●

●

●●

●●

●●

●

●●

●

●

●

●

●

●

●

●

●

●

●

●

●●

●

4

6

8

10

5 7 9pIC50_pred

pIC50

MedChemica | 2016

Future developments

Methodology •  “Metabolophore” extraction •  Rule partition by charge •  Enhanced statistical rule selection methods •  Inferred rule extraction (AB + B C = AC) / matched series / matched networks •  Meta rule identification (eg halogen>>alkyl) •  Rule partition by shape •  Fuzzier atom typing (eg matching indole NH with ArNHC(=O)Me)

Technology •  Transform searching and clustering •  Graph database •  Distributed compute (eg Apache Spark)

Science •  Explaining counter dogma transformations

•  Single crystal x-ray for solubility •  Route of metabolism studies •  Serum protein albumin co-crydtallisation for PPB

•  Cardiac and liver tox screening panel development

78

Bigger, faster, bigger & faster!

MedChemica | 2016

Essentials of the collaboration •  Roche, Genentech and AZ all have ADMET data

processed inside their firewalls to generate transformations (matched pairs transformations with change data)

•  The transformations (fragments of molecules only) are shared with MedChemica

•  MedChemica combines the transformation data and returns the aggregated knowledge

•  Therefore NO party can drill back to anyone else’s structures or original data

•  There is no reach-through by any party •  MedChemica facilitates the science coordinating group

to make suggestions on improvements and enhancements to the data sets, methods for extraction, analysis and exploitation

MedChemica | 2016

How Specific are Pharmacophore dyads?

•  How selective is the pharmacophore? •  What are the odds of it hitting a molecule in the test set vs CHEMBL?

•  Odds of finding in potency set = n(pharmacophore hits in potency set) n(in potency set)

•  Odds of finding in CHEMBL =

n(pharmacophore hits in CHEMBL not in potency set) n(in CHEMBL)

•  Odds ratio = selectivity =

Odds of finding in potency set_______ Odds of finding in CHEMBL(not potency set)

80

27 1470

62 1351211

27/1470 62/1351211

=407 (95% confidence limits: 259-642)

Odds of hitting a potent compound are 407 times greater than a random compound in CHEMBL

Path"

Fragment 1

Fragment 2 [CH2]CN

MedChemica | 2016

Benchmarking Specificity What does a bad odds ratio look like?

What is the odds ratio? Found in CHEMBL 565658/1352681 Found in CHEMBL240 – hERG where pIC50 >=5 1985/2451

OR = 1985/2451 = 0.81 565658/1352681 0.42

=1.94 (95% conf 1.83 – 2.05)

81

Lipophilic base, usually a tertiary amine X = 2-5 atom chain, may include rings, heteroatoms or polar groups

XN

R1

R2

e. g. sertindole: 14nM vs hERG

[$([NX3;H2,H1,H0;!$(N[C,S]=[O,N])]~*~*~*~c),$([NX3;H2,H1,H0;!$(N[C,S]=[O,N])]~*~*~c),$([NX3;H2,H1,H0;!$(N[C,S]=[O,N])]~*~*~*~*~c),$([NX3;H2,H1,H0;!$(N[C,S]=[O,N])]~*~*~*~*~*~c)]

Early simple hERG model

Ar-linker-base has only been found 1.9x more often in hERG inhibitors than at random in ChEMBL

MedChemica | 2016

Fast building block access from CRO collaboration

82

MCExpert suggests

improved building blocks

Specialist synthesis CROs access unique

chemistries

Rapid access to building blocks that address

metabolism and solubility issues

Mono & disubstituted chiral piperidines and pyrollidines

Chiral α methyl aryl amines and alcohols

MedChemica | 2016

Better compounds designed from Data

Essentials Gains

Pains

•  Improved compounds quicker

•  Applicable ideas

•  Confident design decisions

•  Help when stuck

•  Clearly describable plans

•  Maximizing value from ADMET testing

•  Pursuing dead-end series

•  Pursuing dead-end projects

•  Running out of time or $

molecular comparisons by maximum common sub...

Documents